The Perception Gap in Twenty-Seven Years of Data
Perception is shifting — but we can only see it in expensive one-off polls
Source: SF Chamber of Commerce CityBeat 2025 Poll, sponsored by United Airlines
Why the Gap Persists
Total reported crime fell 44.9% from 2019 through 2025 — from 119,177 incidents to 65,707. Every single one of the city's ten police districts recorded a decline. Yet the City Survey hit its lowest safety score since 1996.
The answer lies in composition. Larceny theft — car break-ins, shoplifting, package theft — drives roughly 60% of the statistical decline but is most susceptible to reporting changes. Meanwhile, the visible markers that shape how a street feels — encampments, needles, graffiti, aggressive behavior — are captured in 311 data, not crime data. The gap between what the statistics say and what people experience is the problem Public Safety Pulse solves.
How Does Each Block Feel?
250,000 geocoded 311 disorder reports rendered at block level. Taller, hotter columns = more reports about encampments, street cleaning, graffiti, and needles — a proxy for how unsafe an area feels. Hit play to watch hotspots shift across 24 hours.
What's Driving the Signal?
When Does the City Feel Unsafe?
The same block feels different at noon versus midnight. This is the variation the City Survey can't capture — it asks one question, once every two years. Public Safety Pulse captures it in 4-hour windows, every day, at block level.
Hourly Distribution of 311 Disorder Reports
Disorder vs Crime by Neighborhood
The Disorder–Crime Divergence shows where what people see (311 environmental reports) differs from the crime data. Neighborhoods with high divergence have a perception problem — they feel unsafe because of environmental conditions, not criminal danger. These are where cleaning, lighting, and ambassador programs have the highest ROI.
Monthly Trends
Everything you just saw is proxy data — our best guess.
311 only captures what people report. Crime data only captures what police file. Areas people avoid appear safe in the data. We need the actual signal.
What Phase 1 Unlocks
| What We Have Now (Proxy) | What Phase 1 Adds (Direct) |
|---|---|
| 311 complaints — lagging, reporter bias | Direct, in-the-moment perception |
| Crime incidents — only what gets reported | Real-time safety sentiment |
| Biennial survey — 2-year lag, neighborhood level | Daily signal, block level |
| Review text mining — business-adjacent only | Universal coverage via existing touchpoints |
| Foot traffic proxy — infers avoidance | Directly asks: "How does this area feel?" |
The Solution: Low-Friction Sentiment Capture
A single, optional question — "Right now, how does the surrounding area feel to you?" — delivered through existing digital touchpoints during normal daily activity. Comfortable / Neutral / Uncomfortable. Anonymous. Aggregated by place and time.
The Feedback Loop
Public Safety Pulse creates a powerful feedback loop: collect high-frequency ground truth on how safe people feel → feed it into a correlation engine that identifies which levers move perception → deploy targeted interventions (cleaning, lighting, ambassadors, music, signage) → measure the impact in near-real-time → optimize. This is how you turn data into measurably safer streets.
Partner with us to validate whether direct sentiment can fill the gap.
City Science Lab San Francisco × MIT Media Lab City Science